Research map / Architectures and efficiency / Attention and transformers
Visual Transformers: Where Do Transformers Really Belong in Vision Models?
Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph E. Gonzalez, Kurt Keutzer, Peter Vajda
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use filter-based tokenizer instead of pooling-based (ablation) | VT-ResNet-18 · ImageNet val (VT-ResNet-18) | top-1 accuracy (%): 70.6 -> 72.0 | Improved |
| use filter-based tokenizer instead of pooling-based (ablation) | VT-ResNet-34 · ImageNet val (VT-ResNet-34) | top-1 accuracy (%): 73.5 -> 74.8 | Improved |
| filter-based vs clustering-based tokenizer (ablation) | VT-ResNet · ImageNet val | top-1 accuracy (%): R18 71.5 -> 72.0; R34 75.1 -> 74.8 | Mixed |
| add recurrent tokenizers (ablation) | VT-ResNet-18 · ImageNet val (VT-ResNet-18) | top-1 accuracy (%): 71.4 -> 72.0 | Improved |
| add recurrent tokenizers (ablation) | VT-ResNet-34 · ImageNet val (VT-ResNet-34) | top-1 accuracy (%): 74.5 -> 74.8 | Improved |
| use transformer instead of no token interaction (ablation) | VT-ResNet-18 · ImageNet val (VT-ResNet-18) | top-1 accuracy (%): 68.8 -> 72.0 | Improved |
| use transformer instead of graph convolution (ablation) | VT-ResNet-34 · ImageNet val (VT-ResNet-34) | top-1 accuracy (%): 73.6 -> 74.8 | Improved |
| increase visual tokens from 16 to 64 (ablation) | VT-ResNet-18 · ImageNet val (VT-ResNet-18) | top-1 accuracy (%): 72.0 -> 72.0 | No clear effect |
| increase visual tokens from 16 to 32 (ablation) | VT-ResNet-34 · ImageNet val (VT-ResNet-34) | top-1 accuracy (%): 74.8 -> 75.1 | No clear effect |
| add projector back to feature map (ablation) | VT-ResNet-18 · ImageNet val (VT-ResNet-18) | top-1 accuracy (%): 70.9 -> 72.0 | Improved |
| train with advanced recipe instead of default | VT-ResNet-18 · ImageNet val | top-1 accuracy (%): 72.0 -> 76.8 | Improved |
| replace FPN with VT-FPN | ResNet-50 FPN · COCO-stuff val | mIoU (%): 40.78 -> 41.00 | Improved |
Closest papers by meaning
- Efficient Training of Visual Transformers with Small Datasets (NeurIPS 2021 · similarity 0.92)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · similarity 0.90)
- Intriguing Properties of Vision Transformers (NeurIPS 2021 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
- Vision Transformers: From Semantic Segmentation to Dense Prediction (IJCV 2024 · similarity 0.90)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
